Five-Year Project-Level Statewide Pavement Performance Forecasting Using a Two-Stage Machine Learning Approach Based on Long Short-Term Memory

Abstract

An accurate pavement performance forecasting model is essential for transportation agencies to perform pavement maintenance, rehabilitation, and reconstruction (MR&R) in a predictive and cost-effective manner. Although some forecasting methods have been successful in forecasting short-term (e.g., 1–2 year) pavement conditions at either the project level or network level, accurately forecasting long-term (e.g., 3–5 year) pavement conditions at both project level and network level under real-world conditions is still challenging. Thus, the goal of this paper is to propose a two-stage machine learning approach based on long short-term memory (LSTM) to achieve not only the short-term, but also the long-term, forecasting accuracy at both the project level and network level. The proposed method involves LSTM in the first stage and an artificial neural network (ANN) in the second stage, resulting into a two-stage model. The LSTM first learns the pattern of pavement deterioration based on sequential data (e.g., historical pavement conditions). Then, the ANN further learns the impacts of roadway factors (e.g., traffic parameter, pavement surface type, working district) to adjust the final forecasting results. The accuracy of the proposed two-stage model has been compared with baseline machine learning methods in 2016 on a large, statewide Florida dataset at both the project level and network level to demonstrate the superior capability of the proposed method. In addition, the proposed method has been tested further to forecast future (5-year) pavement conditions (2016–2020). Results show a promising forecasting accuracy for both the short-term and long-term in comparison with the ground truth.

Recurrent Neural Networks for Pavement Performance Forecasting: Review and Model Performance Comparison

“…The two-stage model architecture follows the same functions as its component networks; however, the approaches in Dong et al ( 33 ) and Bukharin et al ( 3 ) slightly differ in the manner of their construction. In the latter, the two-stage network comprises two separate models, while the former utilizes a singular end-to-end model.…”

“…In Dong et al ( 33 ), the LSTM-FCNN delivered a 30.5% RMSE performance improvement over an eXtreme Gradient Boosting regression model in IRI prediction, and outperformed linear regression, FCNN, and gradient boosting decision tree models by an even larger margin. Additionally, in Bukharin et al ( 3 ), another LSTM-FCNN based architecture outperformed a random forest model by an aggregated 18% in deficient-lane-miles error over 5 years and beat an LSTM-only model by 16%.…”

“…For network-level, the primary difficulty is that DOTs may need a 5 year long-term pavement forecast to prepare the budget, with one example being the Florida Department of Transportation (FDOT). FDOT manages over 40,000 lane miles of pavement, and with resurfacing costs averaging over $300,000 per lane mile, a 1% forecasting error at the long-term network level can propagate into a funding gap of over $100 million per year ( 3 ). Comparing this with the $700 to $900 million that FDOT allocates annually to pavement MR&R ( 4 ), a funding disparity on this scale delays treatments enough to have a significant impact on the 3R principle of predictive maintenance.…”

Effects of Different Training Datasets on Machine Learning Models for Pavement Performance Prediction

“…As indicated in previous studies, the most frequently used ML algorithms in pavement performance prediction are support vector regression (SVR), RF, and ANN (8)(9)(10)(11)(12). Therefore, their theoretical bases, characteristics, and calibration hyperparameters are described in the following.…”

“…In addition to the completeness of the database, training data preprocessing has a great impact on a model's predictive performance, is highly dependent on the modeler's experience, and is not typically reported in engineering-related literature (8). In fact, most research focuses on the ML algorithms used and the evaluation metrics, without describing the training database construction and final predictors (8)(9)(10)(11)(12).…”